The day in conversation for 2026-10-23

The discussion: Rosa: That SafeDojo bump—8.25 percentage points on Level I—that’s significant, but it feels like we’re still chasing baseline safety guarantees instead of true robustness in these VLA systems. Dev: I agree, Rosa. The mechanism they introduced, combining the world model imagination with that safety head predicting per-step contact probabilities... it sounds like a clever way to bake conservatism directly into the policy generation rather than just filtering failures post-hoc. It moves from reactive safety to proactive constraint satisfaction during training. Taro: I think it’s interesting because it connects back to our earlier discussions on world model priors we had last week about hierarchical planning. If the world model is already providing a decent structural understanding, the safety head isn't just patching errors; it's intelligently pruning unsafe trajectories before they even start. Rosa: Exactly. But let’s pivot slightly—what about ReStruct? That whole inference-time adaptation idea seems like the real bottleneck for deployment right now if we want these models to actually interact with messy, real-world environments without constant fine-tuning. Dev: ReStruct is the holy grail, but it hinges entirely on how well that task structure skeleton can capture the underlying physics and affordances. If you decompose a complex visuomotor policy into a rigid skeleton, you might lose the necessary nuance needed for subtle interaction preferences unless that decomposition is extremely granular. Taro: That’s my concern. It